fix: improve agent classification, retries, and message delivery

[devhims]

Classification could fail before research started because Fireworks GLM returned incomplete arguments when forced to call classify_request. A subsequent “try again” also lost the original request because failed turns are excluded from completed conversation memory.

This change uses automatic tool selection while requiring exactly one validated classification call before proceeding. Invalid or missing calls receive one repair attempt within the original 20-second deadline, with separate usage accounting, persisted validation diagnostics, and a readable terminal error if repair fails.

The dashboard now renders user messages immediately using React useOptimistic for new sessions and follow-ups. Confirmed admission reconciles the temporary message without duplication. Failed delivery shows an accessible alert and restores the exact draft; ambiguous retries reuse the same idempotency key. Admitted message previews bridge navigation while history loads.

Bare retries now snapshot the failed request's execution message and preserve its original parent. The displayed user message stays unchanged. Resolution is scoped to the account, session, and branch, supports legacy retry chains, and preserves the effective request in subsequent completed conversation memory. An idempotent SQLite migration adds a nullable execution_message column.

Validation

• Platform TypeScript and production Next.js webpack builds pass.
• 31 web unit tests and 16 Chrome dashboard checks pass, including delayed delivery, draft restoration, duplicate prevention, and slow-history navigation.
• 58 focused router, deadline, and conversation-memory tests pass.
• 21 local Workers Durable Object integration tests pass.
• Offline replay of the captured provider failure succeeds after a valid repair through the actual provider adapter and SDK.
• A live Fireworks response missing its clarification question was repaired successfully within 6.3 seconds.

Live classification of the original comparison returned valid output in 2.8 seconds, but GLM still requested clarification rather than selecting topic research. This fixes the reproduced validation failure and retry-context loss; routing quality remains a separate limitation. No full research replay or production deployment was performed.

Test captures and reports remain local and are not included in this PR. No dependency or lockfile changes.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 11, 2026 1:44pm UTC